A Finite-Memory Adaptive Pattern Recognizer

This paper gives an adaptive procedure for selecting a discriminant for a pattern recognizer. The optimum discriminant is selected from a given finite set of discriminants. The selection of this set itself is not considered here. At any stage the optimum discriminant is selected on the basis of the past observations. Since the storage space for these observations is assumed to be limited, and hence the qualifier finite memory, the information stored about these past observations is judiciously selected. No other a priori knowledge is assumed. A mathematical model of the problem of pattern recognition is constructed and several theorems are proved. With the help of these theorems, the adaptive procedure is developed. This adaptive procedure is, in effect, a method of using the finite memory efficiently in "training" the pattern recognizer.